Vertical take-off and landing oriented multi-laser radar low-altitude wind field cooperative inversion and prediction method

Abstract

The application discloses a vertical take-off and landing oriented multi-laser radar low-altitude wind field cooperative inversion and prediction method, relates to the technical field of low-altitude wind field detection and prediction, and comprises the following steps: multi-laser radar cooperative arrangement and joint inversion are performed to obtain a three-dimensional wind field; a multi-working-condition CFD prior database is constructed; a four-channel convolutional neural network is used to realize wind field space reconstruction and turbulence parameter output; a long short-term memory recurrent neural network is used for short-term and nowcasting prediction, and wind speed, wind direction, gust and turbulence parameters in the future 5 minutes to 6 hours are output; after physical consistency correction, the output is combined with VTOL take-off and landing envelope to output a take-off and landing decision result of 'allowance / restriction / prohibition take-off and landing'; and finally visual output is realized. The application realizes the closed-loop fusion of 'physical priori-observation correction-short-term and nowcasting prediction' by using the above-mentioned vertical take-off and landing oriented multi-laser radar low-altitude wind field cooperative inversion and prediction method, and is used for take-off and landing decision and flight path planning of a vertical take-off and landing aircraft.

Description

Technical Field

[0001] This invention relates to the field of low-altitude wind field detection and prediction technology, and in particular to a method for collaborative inversion and prediction of low-altitude wind fields using multiple lidars for vertical take-off and landing. Background Technology

[0002] Vertical takeoff and landing (VTOL) aircraft, including eVTOL, UAS, and other low-altitude air transport vehicles, are extremely sensitive to near-surface low-altitude wind fields during takeoff, hovering, transition, and landing. In such scenarios, wind speed gradients, abrupt changes in wind direction, gusts, turbulence, and local shear caused by building wakes or terrain-based flow all directly affect flight control margins, path tracking accuracy, endurance, and the probability of safe takeoff and landing. Existing research has indicated that wind and turbulence forecasting for advanced air transport has significant operational value, especially the need to output wind speed, wind direction, gusts, and turbulence parameters at high temporal resolution to support go / no-go decisions for flight missions.

[0003] In existing technologies, single-point weather stations can provide local wind speed and direction, but they are unable to reflect the spatial non-uniformity of complex three-dimensional wind fields within the take-off and landing area. Although a single lidar can achieve high-resolution profile observation, it is limited by field-of-view obstruction, scanning geometry, and blind zone issues, making it difficult to independently reconstruct a complete three-dimensional wind field. Existing research shows that dual-Doppler or Doppler lidar combined with synchronous scanning can reconstruct two-dimensional and even three-dimensional wind fields through geometric projection relationships, but requires strict control of synchronization, scanning errors, and uncertainties. Long-range scanning lidar has also been used for quantitative measurement of complex wind fields and upstream flow fields of wind farms.

[0004] On the other hand, while pure computational fluid dynamics methods can provide physically consistent wind field structures, they are sensitive to terrain, buildings, boundary conditions, and turbulence parameters, and have high computational costs, making it difficult to directly meet the minute-level update requirements of low-altitude flight missions. Recent studies have shown that combining high-resolution simulation with machine learning can achieve real-time wind field prediction on complex terrains. For example, a wind field reconstruction method for complex terrain (hereinafter referred to as WindSeer) has been proposed. It trains a convolutional neural network with synthetic data generated by CFD and can predict low-altitude time-averaged wind fields and turbulence intensity in real time on complex terrains, relying only on sparse and noisy measurement data. It can also perform zero-shot generalization on unseen terrains. Another study on short-term wind condition prediction for advanced air traffic (hereinafter referred to as WindAware) shows that, based on high-resolution simulation and sparse ground observations, LSTM-RNN can update wind speed, wind direction, gusts, and turbulence parameters on a 5-minute basis and support short-term nowcast forecasts of up to 6 hours.

[0005] Therefore, the existing technology still has the following problems: First, the three-dimensional wind field observation in the low-altitude take-off and landing area is incomplete; second, although multiple lidars can improve the observation coverage, there is a lack of a unified joint inversion and quality control mechanism; third, CFD and machine learning are often used separately, lacking a closed-loop fusion of "physical prior - observation correction - short-term prediction"; fourth, the output of wind shear, gust and turbulence risk for VTOL take-off and landing envelope is not detailed enough. Summary of the Invention

[0006] The purpose of this invention is to provide a method for collaborative inversion and prediction of low-altitude wind fields using multiple lidars for vertical take-off and landing, in order to solve the problems of sparse low-altitude wind field observations, incomplete three-dimensional reconstruction, insufficient real-time performance of numerical simulation, and lack of physical constraints in machine learning in the existing technology.

[0007] To achieve the above objectives, this invention provides a method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing, comprising: S1. At least two ground-based Doppler lidars with multi-view cross-observation capabilities are deployed around the VTOL take-off and landing field and its arrival and departure corridors, depending on the environment and wind direction. Vehicle-mounted or airborne lidars are also deployed in the blind zone to obtain wind field data with high spatiotemporal resolution. S2. Synchronize, unify coordinates and control the quality of data from multiple lidar and auxiliary weather stations, filter out outliers and fill in missing values ​​to form a consistent spatiotemporal sampling grid. S3. Based on multi-radar spatiotemporal synchronization and quality control, the radial velocity is projected onto the voxel grid, and the three-dimensional wind vector field is jointly inverted through weighted least squares and regularization constraints, and the local missing regions are filled in. S4. Based on the terrain and building data around the take-off and landing field, a three-dimensional computational domain is established. A prior wind field database under different combinations of wind direction, wind speed and atmospheric stability is generated through multi-condition CFD simulation, which serves as the training set and physical prior for the machine learning model. S5. Construct a four-channel convolutional neural network, using the observation mask, terrain distance field, wind vector components and uncertainty as inputs, and fine-tune it with CFD prior training and measured data to realize the spatial reconstruction of the three-dimensional wind field and the output of turbulence parameters. S6. Based on the spatial reconstruction results of the wind field, construct a time series prediction model based on LSTM-RNN, input historical wind field voxel sequences and auxiliary data, and output predicted values ​​of wind speed, wind direction, gusts and turbulence parameters for the next 5 to 6 hours. S7. Apply physical consistency corrections such as mass conservation, boundary constraints, and time smoothing to the short-term prediction results, and set risk thresholds in conjunction with the VTOL takeoff and landing envelope to output the judgment result of takeoff and landing feasibility. S8. The final prediction and assessment results will be visualized and output in the form of three-dimensional wind field map, risk heat map, profile map and alarm information, so that the flight control or ground command system can use them for take-off and landing decisions and trajectory planning.

[0008] Preferably, step S1 specifically includes: Determine the scope of the target vertical takeoff and landing (VTOL) aircraft's takeoff and landing field and its approach and departure corridors, and analyze the terrain undulations, building distribution, obstacle locations, and prevailing wind direction; At least two ground-based scanning Doppler lidars, preferably three or more, are deployed around the take-off and landing field and its arrival and departure corridors, located on the upwind, lateral and downwind sides of the take-off and landing field respectively, to form a multi-view cross observation geometry; A vehicle-mounted or airborne lidar will be deployed in the low-altitude blind zone near the ground to fill in the blind zone data. Control each lidar to perform planar position indication scanning, distance and height indication scanning, or virtual wind tower scanning mode to acquire radial wind speed, scanning azimuth angle, elevation angle, distance gate information, and signal-to-noise ratio information.

[0009] Preferably, step S2 specifically includes: The observation data from each lidar and auxiliary meteorological station are uniformly timestamped, and Coordinated Universal Time is used as the global time reference. Coordinate transformation is performed on the attitude angle, installation angle, azimuth angle and scanning geometry of different devices to map the output of each sensor to a unified Cartesian coordinate system; Outlier removal, low signal-to-noise ratio echo filtering, short-time missing value interpolation, and scan inconsistency correction are performed on radial wind speed. For samples with scanning time sequence deviations, a combination of linear time interpolation and local weighted smoothing is used to restore a consistent spatiotemporal sampling grid.

[0010] Preferably, the specific content of step S3 is as follows: The radial velocity observations obtained by multiple lidars at the same time are uniformly projected onto the target voxel grid to establish the observation equation between radial velocity and three-dimensional wind vector; For the wind vector components within each voxel We employ a weighted least squares method to solve the problem, and add Tikhonov regularization terms or spatial smoothing constraints to suppress underdetermined problems and observation noise propagation. The weights of each observation are determined by combining the signal-to-noise ratio, ranging error, scanning incident angle, obstruction probability, and inter-radar geometry. For local areas not covered by joint observation from multiple radars, the flow continuity constraints and short-time evolution constraints of adjacent voxels are used to fill the gaps. Output a three-dimensional initial wind field that matches the actual takeoff and landing field spatial range, including horizontal wind speed, vertical velocity, wind direction, wind shear, and local turbulence indices.

[0011] Preferably, the specific content of step S4 is as follows: A three-dimensional computational domain is established based on the digital elevation model, building model and surface roughness distribution around the take-off and landing field. A multi-condition computational fluid dynamics database is generated based on different combinations of prevailing wind direction, wind speed level, and atmospheric stability. The steady-state average wind field, solved by Reynolds-averaged Navier-Stokes method, is used as the background field. Local high-resolution subdomains are used to refine the solution for the take-off and landing platform, building wake region, obstacle edge, and strong shear region near the ground. Nested large eddy simulations or local transient simulations may be used when necessary to obtain fine-grained samples of wind speed, wind direction, and turbulence. The computational fluid dynamics database was used as the training set for subsequent machine learning models and as the physical prior for joint inversion and short-term prediction.

[0012] Preferably, the specific content of step S5 is as follows: A convolutional neural network for three-dimensional wind field spatial reconstruction was constructed, using a four-channel input structure; The system is configured with four input channels: the first channel is the observation mask, used to characterize whether the voxel has a real measurement; the second channel is the terrain distance field or terrain coding field, used to express the geometric constraints of the take-off and landing area and surrounding obstacles on the flow field; the third channel is the horizontal wind speed or wind vector component; and the fourth channel is the uncertainty or confidence level of the radar-inverted wind field. The model output is set to a three-dimensional wind field with the same resolution as the input space. Simultaneously output turbulent kinetic energy or local gust intensity; Pre-training of convolutional neural networks using computational fluid dynamics-synthesized samples; The measured wind field obtained by joint inversion of multiple lidars is used to fine-tune the pre-trained network, forming a closed loop of computational fluid dynamics priors, observation correction and spatial reconstruction. The measured data is input into the trained network, which outputs the three-dimensional wind field reconstruction results and turbulence parameters.

[0013] Preferably, the specific content of step S6 is as follows: Based on the results of three-dimensional wind field spatial reconstruction, a short-term prediction model is constructed, and a long short-term memory recurrent neural network is used as the time series backbone network. The three-dimensional wind field features at several consecutive time points are encoded and input into a long short-term memory recurrent neural network. The wind direction is decomposed into horizontal orthogonal components, and sequential modeling is performed on each component. The wind direction angle is then reconstructed by vector synthesis after prediction. Historical wind field voxel sequences, takeoff and landing field ground meteorological station data, radar location information, terrain coding, and current flight window status are used as input features; Output predicted values ​​for wind speed, wind direction, gusts, and turbulence parameters for the next 5, 10, 15 minutes, and up to 6 hours.

[0014] Preferably, the specific content of step S7 is as follows: Perform physical consistency correction on the short-term prediction results output in step S6; Mass conservation or approximately incompressible constraints are applied to the predicted wind field to suppress local unreasonable wind vector jumps; Apply slope boundary or no-slip corrections to areas near terrain and building wakes to prevent prediction results from penetrating solid boundaries; Smoothing constraints are applied to the time evolution of wind speed, wind direction, and turbulence parameters to reduce short-term jitter; Based on the takeoff and landing envelope of the vertical takeoff and landing aircraft, risk thresholds are set for the horizontal crosswind component, vertical velocity, wind shear rate, gust factor and turbulence intensity; Based on the comparison between the revised prediction results and the risk threshold, the system outputs a judgment result of "allow take-off and landing", "restrict take-off and landing", or "prohibit take-off and landing".

[0015] Preferably, the specific content of step S8 is as follows: The final output will be visualized in the form of a 3D wind field map, a risk heat map, a profile map, a wind condition summary for the take-off and landing window, and alarm information. The visualization results should include at least the wind speed vector field, wind shear distribution, turbulence intensity distribution, and risk level zoning map for the current moment and a short future moment; The visualization results are output to the flight control system or ground command system for mission planning, trajectory correction, and takeoff and landing decisions for UAVs, electric vertical takeoff and landing vehicles, and other vertical takeoff and landing platforms.

[0016] Therefore, the present invention employs the above-mentioned multi-lidar low-altitude wind field collaborative inversion and prediction method for vertical take-off and landing, which has the following beneficial effects: (1) By coordinating the deployment and joint inversion of multiple lidars, the observation coverage of low-altitude take-off and landing areas can be significantly improved, reducing the problems of limited field of view and numerous blind spots of a single lidar, thereby obtaining a more complete three-dimensional wind field structure. Doppler lidar and scanning lidar have been proven to be suitable for quantitative reconstruction of complex flow fields, so this invention can achieve reliable space wind field recovery in engineering.

[0017] (2) Using computational fluid dynamics as a physical prior and convolutional neural networks as a wind field spatial reconstructor can significantly improve prediction efficiency while maintaining physical consistency. Existing research has shown that convolutional neural networks trained on computational fluid dynamics synthetic data can recover low-altitude wind fields in real time using sparse noise observations on complex terrain and have zero-sample generalization ability, which provides a direct technical basis for the "computational fluid dynamics + convolutional neural network" approach of this invention.

[0018] (3) Using a long short-term memory recurrent neural network for short-term prediction can learn the temporal patterns from sparse ground observations, joint inversion results, and computational fluid dynamics priors, and output wind speed, wind direction, gusts, and turbulence parameters that are truly useful for the take-off and landing of vertical take-off and landing aircraft, rather than just giving a general average wind speed. Existing research has shown that this approach is suitable for minute-level updates and forecasts up to 6 hours in advanced air traffic scenarios.

[0019] (4) It can output risk classification results for take-off and landing scenarios of vertical take-off and landing aircraft, and directly map the physical quantities of wind shear, gusts and turbulence into take-off and landing judgment information of "take-off and landing allowed", "take-off and landing restricted" or "take-off and landing prohibited", which can be quickly called by the ground system or flight control system. This output method is consistent with the engineering requirements of advanced air traffic for take-off and landing decisions.

[0020] (5) It has good scalability: it can operate using only ground-based Doppler lidar, or it can be expanded into a multi-source system of "ground-based radar + airborne radar or UAV supplementary measurement + ground meteorological station + computational fluid dynamics prior + machine learning", which is suitable for different scenarios such as fixed vertical take-off and landing fields, temporary take-off and landing points and complex urban canyons.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an embodiment of the multi-lidar low-altitude wind field collaborative inversion and prediction method for vertical take-off and landing according to the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0025] Example Please see Figure 1 This invention provides a method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing, including: S1. At least two ground-based Doppler lidars with multi-view cross-observation capabilities are deployed around the VTOL take-off and landing field and its approach and departure corridors, depending on the environment and wind direction. Vehicle-mounted or airborne lidars are also deployed in blind areas to obtain wind field data with high spatiotemporal resolution.

[0026] Around the target VTOL takeoff and landing field and its arrival and departure corridors, at least two ground-based scanning Doppler lidars are deployed, preferably three or more, to form a multi-view cross-observation geometry, taking into account terrain undulations, building distribution, obstacle distribution, and prevailing wind direction. Each lidar is deployed on the upwind, lateral, and leeward sides of the takeoff and landing field. If necessary, a vehicle-mounted or airborne lidar is added at blind spots to fill low-altitude near-surface blind areas. Each lidar performs PPI scanning, RHI scanning, or virtual anemometer tower scanning modes to acquire radial wind speed, scanning azimuth, elevation angle, range gate information, and signal-to-noise ratio information. This deployment method draws on the mature application of Doppler anemometers and scanning lidars in complex flow fields, while also responding to the demand for high spatiotemporal resolution wind field input in advanced air traffic.

[0027] S2. Time synchronization, coordinate unification and quality control are performed on data from multiple lidar and auxiliary weather stations. Outliers are filtered out and missing values ​​are filled to form a consistent spatiotemporal sampling grid.

[0028] The observation data from various lidar and auxiliary meteorological stations were standardized with unified timestamps, using UTC as the global time reference. Coordinate transformations were performed on the attitude angles, installation angles, azimuth angles, and scanning geometry of different devices, mapping the outputs of each sensor to a unified Cartesian coordinate system. Subsequently, outlier removal, low signal-to-noise ratio echo filtering, short-term missing data interpolation, and scanning inconsistency correction were applied to radial wind speed. For samples with inter-scan temporal deviations, a combination of linear time interpolation and local weighted smoothing was used to restore a consistent spatiotemporal sampling grid. Uncertainty control and statistical smoothing of multi-scan lidar are key aspects repeatedly emphasized in existing wind field measurement studies.

[0029] S3. Based on multi-radar spatiotemporal synchronization and quality control, the radial velocity is projected onto the voxel grid, and the three-dimensional wind vector field is jointly inverted through weighted least squares and regularization constraints, and the missing local regions are filled in.

[0030] After synchronization and quality control, radial velocity observations obtained from multiple lidars at the same time are projected onto the target voxel grid, establishing the observation equation between radial velocity and three-dimensional wind vector. For each voxel, the wind vector components... A weighted least squares method is employed for solution, incorporating a Tikhonov regularization term or spatial smoothing constraint to suppress underdetermined problems and observation noise propagation. The weights are determined by the signal-to-noise ratio, ranging error, scanning incident angle, obstruction probability, and inter-radar geometry. For local areas not covered by multi-radar observations, flow continuity constraints and short-time evolution constraints of adjacent voxels are used for completion. This step generates a three-dimensional initial wind field consistent with the actual takeoff and landing field spatial range, including horizontal wind speed, vertical velocity, wind direction, wind shear, and local turbulence indices. Dual-Doppler and Doppler lidar (Light Detection and Ranging) have been proven effective for wind field reconstruction, while simultaneous scanning, error analysis, and uncertainty control are prerequisites for ensuring the reliability of quantitative results.

[0031] S4. Based on the terrain and building data around the take-off and landing field, a three-dimensional computational domain is established. A prior wind field database under different combinations of wind direction, wind speed and atmospheric stability is generated through multi-condition CFD simulation, which serves as the training set and physical prior for the machine learning model.

[0032] A three-dimensional computational domain is established based on the digital elevation model, building model, and surface roughness distribution around the takeoff and landing field. A multi-condition CFD database is generated according to combinations of prevailing wind direction, wind speed level, and atmospheric stability. Preferably, the background field is solved using RANS (Reynolds-Averaged Navier-Stokes equations) to obtain the steady-state average wind field. Key areas such as the takeoff and landing platform, building wake zone, obstacle edges, and near-surface strong shear zone are solved using local high-resolution subdomains. Nested LES or local transient simulations are used when necessary to obtain finer-grained wind speed, wind direction, and turbulence samples. This CFD database serves as both a training set for machine learning models and a physical prior for subsequent joint inversion and short-term prediction. In complex terrain and urban / low-altitude flight scenarios, CFD remains an important source for generating high-resolution wind fields, but its drawbacks include sensitivity to boundary conditions and insufficient real-time performance. Therefore, it is more suitable as a teacher or prior for data-driven models.

[0033] S5. Construct a four-channel convolutional neural network, using the observation mask, terrain distance field, wind vector components and uncertainty as inputs. Through CFD prior training and fine-tuning with measured data, realize the spatial reconstruction of the three-dimensional wind field and the output of turbulence parameters.

[0034] After obtaining the joint inversion results from multiple lidar sensors and the CFD prior, a convolutional neural network for three-dimensional wind field spatial reconstruction is constructed. Preferably, a four-channel input structure isomorphic to WindSeer is adopted, with the four input channels as follows: the first channel is the observation mask, used to characterize which voxels have real measurements; the second channel is the terrain distance field or terrain coding field, used to express the geometric constraints of the take-off and landing area and surrounding obstacles on the flow field; the third channel is the horizontal wind speed or wind vector component channel; and the fourth channel is the uncertainty or confidence level channel of the radar-inverted wind field. The model output is a three-dimensional wind field with the same resolution as the input space. It can simultaneously output turbulent kinetic energy (TKE) or local gust intensity. The network is pre-trained using CFD synthetic samples and then fine-tuned using measured radar joint inversion fields, thus achieving a closed loop of "CFD prior—observation correction—spatial reconstruction." WindSeer has demonstrated that convolutional neural networks trained on CFD synthetic data can reconstruct low-altitude wind fields in real time over complex terrain using only sparse, noisy measurements, and output three-dimensional wind components and turbulence intensity.

[0035] S6. Based on the spatial reconstruction results of the wind field, construct a time series prediction model based on LSTM-RNN. Input historical wind field voxel sequences and auxiliary data, and output predicted values ​​of wind speed, wind direction, gusts and turbulence parameters for the next 5 to 6 hours.

[0036] Based on the spatial reconstruction results, a short-term prediction model is further constructed. Preferably, an LSTM-RNN is used as the time series backbone network to encode the three-dimensional wind field features at several consecutive time points, outputting predicted values ​​for wind speed, wind direction, gusts, and turbulence parameters for the next 5, 10, 15 minutes, or even 6 hours. For wind direction prediction, the wind direction can be decomposed into horizontal orthogonal components. and Sequence modeling is performed separately, and the wind direction angle is reconstructed through vector synthesis after prediction. The input features include not only historical wind field voxel sequences, but also takeoff and landing field ground meteorological station data, radar confidence, terrain coding, and current flight window status. This LSTM-RNN structure is adopted because existing research on advanced air traffic has shown that LSTM-RNN can utilize sparse ground observations and high-resolution simulation data to perform high-frequency nowcasts of wind speed, wind direction, gusts, and turbulence parameters, and can support the safe operation of UAS / eVTOL.

[0037] S7. Apply physical consistency corrections such as mass conservation, boundary constraints, and time smoothing to the short-term prediction results, and set risk thresholds in conjunction with the VTOL takeoff and landing envelope to output the judgment result of takeoff and landing feasibility.

[0038] The short-term prediction results output in step S6 are corrected for physical consistency, mainly in three aspects: First, mass conservation or approximately incompressible constraints are applied to the predicted wind field to suppress unreasonable local wind vector jumps; second, slope boundary or no-slip corrections are applied to areas near terrain and building wakes to prevent prediction results from penetrating solid boundaries; third, smoothing constraints are set on the temporal evolution of wind speed, wind direction, and turbulence parameters to reduce short-term jitter. Subsequently, risk thresholds are set for horizontal crosswind components, vertical velocity, wind shear rate, gust factor, and turbulence intensity based on the VTOL takeoff and landing envelope, and the judgment results of "allow takeoff and landing," "restricted takeoff and landing," or "prohibited takeoff and landing" are output. For AAM scenarios, the nowcast results of wind and turbulence parameters can be directly used for flight path avoidance and go / no-go decisions, which has been clearly verified in WindAware.

[0039] S8. The final prediction and assessment results will be visualized and output in the form of three-dimensional wind field map, risk heat map, profile map and alarm information, so that the flight control or ground command system can use them for take-off and landing decisions and trajectory planning.

[0040] The final output will be presented to the flight control system or ground command system in the form of a 3D wind field map, risk heat map, profile map, wind condition summary for takeoff and landing windows, and alarm information. The visualization results should include at least the wind speed vector field, wind shear distribution, turbulence intensity distribution, and risk level zoning map for the current moment and short future moments, enabling operators to quickly identify suitable takeoff and landing windows and high-risk areas. For UAVs, eVTOL, and other VTOL platforms, this result can be directly used for mission planning, trajectory correction, and takeoff and landing decisions.

[0041] Therefore, the present invention adopts the above-mentioned multi-lidar low-altitude wind field collaborative inversion and prediction method for vertical take-off and landing, which can form a multi-station, multi-view, and multi-scale collaborative observation network around the take-off and landing area, and on this basis, construct an integrated technical link for three-dimensional wind field reconstruction, short-term prediction and take-off and landing risk assessment.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing, characterized in that, include: S1. At least two ground-based Doppler lidars with multi-view cross-observation capabilities are deployed around the VTOL take-off and landing field and its arrival and departure corridors, depending on the environment and wind direction. Vehicle-mounted or airborne lidars are also deployed in the blind zone to obtain wind field data with high spatiotemporal resolution. S2. Synchronize, unify coordinates and control the quality of data from multiple lidar and auxiliary weather stations, filter out outliers and fill in missing values ​​to form a consistent spatiotemporal sampling grid. S3. Based on multi-radar spatiotemporal synchronization and quality control, the radial velocity is projected onto the voxel grid, and the three-dimensional wind vector field is jointly inverted through weighted least squares and regularization constraints, and the local missing regions are filled in. S4. Based on the terrain and building data around the take-off and landing field, a three-dimensional computational domain is established. A prior wind field database under different combinations of wind direction, wind speed and atmospheric stability is generated through multi-condition CFD simulation, which serves as the training set and physical prior for the machine learning model. S5. Construct a four-channel convolutional neural network, using the observation mask, terrain distance field, wind vector components and uncertainty as inputs, and fine-tune it with CFD prior training and measured data to realize the spatial reconstruction of the three-dimensional wind field and the output of turbulence parameters. S6. Based on the spatial reconstruction results of the wind field, construct a time series prediction model based on LSTM-RNN, input historical wind field voxel sequences and auxiliary data, and output predicted values ​​of wind speed, wind direction, gusts and turbulence parameters for the next 5 to 6 hours. S7. Apply physical consistency corrections such as mass conservation, boundary constraints, and time smoothing to the short-term prediction results, and set risk thresholds in conjunction with the VTOL takeoff and landing envelope to output the judgment result of takeoff and landing feasibility. S8. The final prediction and assessment results will be visualized and output in the form of three-dimensional wind field map, risk heat map, profile map and alarm information, so that the flight control or ground command system can use them for take-off and landing decisions and trajectory planning.

2. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 1, characterized in that, The specific content of step S1 is as follows: Determine the scope of the target vertical takeoff and landing (VTOL) aircraft's takeoff and landing field and its approach and departure corridors, and analyze the terrain undulations, building distribution, obstacle locations, and prevailing wind direction; At least two ground-based scanning Doppler lidars are deployed around the take-off and landing field and its arrival and departure corridors, located on the upwind, lateral, and downwind sides of the take-off and landing field, respectively, to form a multi-view cross-observation geometry. A vehicle-mounted or airborne lidar will be deployed in the low-altitude blind zone near the ground to fill in the blind zone data. Control each lidar to perform planar position indication scanning, distance and height indication scanning, or virtual wind tower scanning mode to acquire radial wind speed, scanning azimuth angle, elevation angle, distance gate information, and signal-to-noise ratio information.

3. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 2, is characterized in that... The specific content of step S2 is as follows: The observation data from each lidar and auxiliary meteorological station are uniformly timestamped, and Coordinated Universal Time is used as the global time reference. Coordinate transformation is performed on the attitude angle, installation angle, azimuth angle and scanning geometry of different devices to map the output of each sensor to a unified Cartesian coordinate system; Outlier removal, low signal-to-noise ratio echo filtering, short-time missing value interpolation, and scan inconsistency correction are performed on radial wind speed. For samples with scanning time sequence deviations, a combination of linear time interpolation and local weighted smoothing is used to restore a consistent spatiotemporal sampling grid.

4. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 3, is characterized in that... The specific content of step S3 is as follows: The radial velocity observations obtained by multiple lidars at the same time are uniformly projected onto the target voxel grid to establish the observation equation between radial velocity and three-dimensional wind vector; For the wind vector components within each voxel The weighted least squares method is used to solve the problem, and Tikhonov regularization or spatial smoothing constraints are added. The weights of each observation are determined by combining the signal-to-noise ratio, ranging error, scanning incident angle, obstruction probability, and inter-radar geometry. For local areas not covered by joint observation from multiple radars, the flow continuity constraints and short-time evolution constraints of adjacent voxels are used to fill the gaps. Output a three-dimensional initial wind field that matches the actual takeoff and landing field spatial range, including horizontal wind speed, vertical velocity, wind direction, wind shear, and local turbulence indices.

5. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 4, characterized in that, The specific content of step S4 is as follows: A three-dimensional computational domain is established based on the digital elevation model, building model and surface roughness distribution around the take-off and landing field. A multi-condition computational fluid dynamics database is generated based on different combinations of prevailing wind direction, wind speed level, and atmospheric stability. The steady-state average wind field, solved by Reynolds-averaged Navier-Stokes method, is used as the background field. Local high-resolution subdomains are used to refine the solution for the take-off and landing platform, building wake region, obstacle edge, and strong shear region near the ground. Nested large eddy simulations or local transient simulations may be used when necessary to obtain fine-grained samples of wind speed, wind direction, and turbulence. The computational fluid dynamics database is used as the training set for subsequent machine learning models and as the physical prior for joint inversion and short-term prediction.

6. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 5, is characterized in that... The specific content of step S5 is as follows: A convolutional neural network for three-dimensional wind field spatial reconstruction was constructed, using a four-channel input structure; The system is configured with four input channels: the first channel is the observation mask, used to characterize whether the voxel has a real measurement; the second channel is the terrain distance field or terrain coding field, used to express the geometric constraints of the take-off and landing area and surrounding obstacles on the flow field; the third channel is the horizontal wind speed or wind vector component; and the fourth channel is the uncertainty or confidence level of the radar-inverted wind field. The model output is set to a three-dimensional wind field with the same resolution as the input space. Simultaneously output turbulent kinetic energy or local gust intensity; Pre-training of convolutional neural networks using computational fluid dynamics-synthesized samples; The measured wind field obtained by joint inversion of multiple lidars is used to fine-tune the pre-trained network, forming a closed loop of computational fluid dynamics priors, observation correction and spatial reconstruction. The measured data is input into the trained network, which outputs the three-dimensional wind field reconstruction results and turbulence parameters.

7. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 6, is characterized in that... The specific content of step S6 is as follows: Based on the results of three-dimensional wind field spatial reconstruction, a short-term prediction model is constructed, and a long short-term memory recurrent neural network is used as the time series backbone network. The three-dimensional wind field features at several consecutive time points are encoded and input into a long short-term memory recurrent neural network. Decompose wind direction into horizontal orthogonal components. and Sequence modeling is performed separately, and the wind direction angle is reconstructed by vector synthesis after prediction. Historical wind field voxel sequences, takeoff and landing field ground meteorological station data, radar location information, terrain coding, and current flight window status are used as input features; Output predicted values ​​for wind speed, wind direction, gusts, and turbulence parameters for the next 5, 10, 15 minutes, and up to 6 hours.

8. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 7, is characterized in that... The specific content of step S7 is as follows: Perform physical consistency correction on the short-term prediction results output in step S6; Mass conservation or approximately incompressible constraints are applied to the predicted wind field to suppress local unreasonable wind vector jumps; Apply slope boundary or no-slip corrections to areas near terrain and building wakes to prevent prediction results from penetrating solid boundaries; Smoothing constraints are applied to the time evolution of wind speed, wind direction, and turbulence parameters to reduce short-term jitter; Based on the takeoff and landing envelope of the vertical takeoff and landing aircraft, risk thresholds are set for the horizontal crosswind component, vertical velocity, wind shear rate, gust factor and turbulence intensity; Based on the comparison between the revised prediction results and the risk threshold, the system outputs a judgment result of "allow take-off and landing", "restrict take-off and landing", or "prohibit take-off and landing".

9. The method for collaborative inversion and prediction of low-altitude wind fields using multiple lidar sensors for vertical takeoff and landing as described in claim 8, characterized in that, The specific content of step S8 is as follows: The final output will be visualized in the form of a 3D wind field map, a risk heat map, a profile map, a wind condition summary for the take-off and landing window, and alarm information. The visualization results should include at least the wind speed vector field, wind shear distribution, turbulence intensity distribution, and risk level zoning map for the current moment and a short future moment; The visualization results are output to the flight control system or ground command system for mission planning, trajectory correction, and takeoff and landing decisions for UAVs, electric vertical takeoff and landing vehicles, and other vertical takeoff and landing platforms.

Images